}
glm.ss_type = lex_integer (lexer);
- if (1 != glm.ss_type && 2 != glm.ss_type )
+ if (1 > glm.ss_type && 3 < glm.ss_type )
{
- msg (ME, _("Only types 1 & 2 sum of squares are currently implemented"));
+ msg (ME, _("Only types 1, 2 & 3 sums of squares are currently implemented"));
goto error;
}
gsl_matrix_free (cm);
}
+/*
+ Type 3 sums of squares.
+ Populate SSQ with the Type 2 sums of squares according to COV
+ */
+static void
+ssq_type3 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd)
+{
+ gsl_matrix *cm = covariance_calculate_unnormalized (cov);
+ size_t i;
+ size_t k;
+ bool *model_dropped = xcalloc (covariance_dim (cov), sizeof (*model_dropped));
+ bool *submodel_dropped = xcalloc (covariance_dim (cov), sizeof (*submodel_dropped));
+ const struct categoricals *cats = covariance_get_categoricals (cov);
+
+ double ss0;
+ gsl_matrix *submodel_cov = gsl_matrix_alloc (cm->size1, cm->size2);
+ fill_submatrix (cm, submodel_cov, submodel_dropped);
+ reg_sweep (submodel_cov, 0);
+ ss0 = gsl_matrix_get (submodel_cov, 0, 0);
+ gsl_matrix_free (submodel_cov);
+ free (submodel_dropped);
+
+ for (k = 0; k < cmd->n_interactions; k++)
+ {
+ gsl_matrix *model_cov = NULL;
+ size_t n_dropped_model = 0;
+
+ for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++)
+ {
+ const struct interaction * x =
+ categoricals_get_interaction_by_subscript (cats, i - cmd->n_dep_vars);
+
+ model_dropped[i] = false;
+
+ if ( cmd->interactions [k] == x)
+ {
+ assert (n_dropped_model < covariance_dim (cov));
+ n_dropped_model++;
+ model_dropped[i] = true;
+ }
+ }
+
+ model_cov = gsl_matrix_alloc (cm->size1 - n_dropped_model, cm->size2 - n_dropped_model);
+
+ fill_submatrix (cm, model_cov, model_dropped);
+
+ reg_sweep (model_cov, 0);
+
+ gsl_vector_set (ssq, k + 1,
+ gsl_matrix_get (model_cov, 0, 0) - ss0);
+
+ gsl_matrix_free (model_cov);
+ }
+ free (model_dropped);
+
+ gsl_matrix_free (cm);
+}
+
+
+
//static void dump_matrix (const gsl_matrix *m);
static void
ssq_type1 (cov, ws.ssq, cmd);
break;
case 2:
- case 3:
- /* Type 3 is not yet implemented :( but for balanced designs it is the same as type 2 */
ssq_type2 (cov, ws.ssq, cmd);
break;
+ case 3:
+ ssq_type3 (cov, ws.ssq, cmd);
+ break;
default:
NOT_REACHED ();
break;
const struct fmt_spec *wfmt =
cmd->wv ? var_get_print_format (cmd->wv) : &F_8_0;
+ double intercept_ssq;
+ double ssq_effects;
double n_total, mean;
- double df_corr = 0.0;
+ double df_corr = 1.0;
double mse = 0;
int f;
struct tab_table *t;
const int nc = 6;
- int nr = heading_rows + 4 + cmd->n_interactions;
+ int nr = heading_rows + 3 + cmd->n_interactions;
if (cmd->intercept)
- nr++;
+ nr += 2;
msg (MW, "GLM is experimental. Do not rely on these results.");
t = tab_create (nc, nr);
moments_calculate (ws->totals, &n_total, &mean, NULL, NULL, NULL);
- if (cmd->intercept)
- df_corr += 1.0;
-
df_corr += categoricals_df_total (ws->cats);
- mse = gsl_vector_get (ws->ssq, 0) / (n_total - df_corr);
-
r = heading_rows;
- tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Model"));
+ if (cmd->intercept)
+ tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Model"));
+ else
+ tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Model"));
r++;
+ mse = gsl_vector_get (ws->ssq, 0) / (n_total - df_corr);
+
+ intercept_ssq = pow2 (mean * n_total) / n_total;
+
+ ssq_effects = 0.0;
if (cmd->intercept)
{
- const double intercept = pow2 (mean * n_total) / n_total;
const double df = 1.0;
- const double F = intercept / df / mse;
+ const double F = intercept_ssq / df / mse;
tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Intercept"));
- tab_double (t, 1, r, 0, intercept, NULL);
+ tab_double (t, 1, r, 0, intercept_ssq, NULL);
tab_double (t, 2, r, 0, 1.00, wfmt);
- tab_double (t, 3, r, 0, intercept / df, NULL);
+ tab_double (t, 3, r, 0, intercept_ssq / df, NULL);
tab_double (t, 4, r, 0, F, NULL);
tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr),
NULL);
for (f = 0; f < cmd->n_interactions; ++f)
{
struct string str = DS_EMPTY_INITIALIZER;
- const double df = categoricals_df (ws->cats, f);
- const double ssq = gsl_vector_get (ws->ssq, f + 1);
- const double F = ssq / df / mse;
+ double df = categoricals_df (ws->cats, f);
+
+ double ssq = gsl_vector_get (ws->ssq, f + 1);
+ double F;
+
+ ssq_effects += ssq;
+
+ if (! cmd->intercept)
+ {
+ df++;
+ ssq += intercept_ssq;
+ }
+
+ F = ssq / df / mse;
interaction_to_string (cmd->interactions[f], &str);
tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, ds_cstr (&str));
ds_destroy (&str);
}
{
- /* Corrected Model */
- const double df = df_corr - 1.0;
- const double ssq = ws->total_ssq - gsl_vector_get (ws->ssq, 0);
- const double F = ssq / df / mse;
+ /* Model / Corrected Model */
+ double df = df_corr;
+ double ssq = ws->total_ssq - gsl_vector_get (ws->ssq, 0);
+ double F;
+
+ if ( cmd->intercept )
+ df --;
+ else
+ ssq += intercept_ssq;
+
+ F = ssq / df / mse;
tab_double (t, 1, heading_rows, 0, ssq, NULL);
tab_double (t, 2, heading_rows, 0, df, wfmt);
tab_double (t, 3, heading_rows, 0, ssq / df, NULL);
tab_double (t, 3, r++, 0, mse, NULL);
}
+ {
+ tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Total"));
+ tab_double (t, 1, r, 0, ws->total_ssq + intercept_ssq, NULL);
+ tab_double (t, 2, r, 0, n_total, wfmt);
+
+ r++;
+ }
+
if (cmd->intercept)
{
- const double intercept = pow2 (mean * n_total) / n_total;
- const double ssq = intercept + ws->total_ssq;
-
- tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Total"));
- tab_double (t, 1, r, 0, ssq, NULL);
- tab_double (t, 2, r, 0, n_total, wfmt);
-
- r++;
+ tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Total"));
+ tab_double (t, 1, r, 0, ws->total_ssq, NULL);
+ tab_double (t, 2, r, 0, n_total - 1.0, wfmt);
}
- tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Total"));
-
-
- tab_double (t, 1, r, 0, ws->total_ssq, NULL);
- tab_double (t, 2, r, 0, n_total - 1.0, wfmt);
-
tab_submit (t);
}